Continual learning is a long-standing challenge in robot policy learning, where a policy must acquire new skills over time without catastrophically forgetting previously learned ones. While prior work has extensively studied continual learning in relatively small behavior cloning (BC) policy models trained from scratch, its behavior in modern large-scale pretrained Vision-Language-Action (VLA) models remains underexplored. In this work, we found that pretrained VLAs are remarkably resistant to forgetting compared with smaller policy models trained from scratch. Simple Experience Replay (ER) works surprisingly well on VLAs, sometimes achieving zero forgetting even with a small replay data size. Our analysis reveals that pretraining plays a critical role in downstream continual learning performance: large pretrained models mitigate forgetting with a small replay buffer size while maintaining strong forward learning capabilities. Furthermore, we found that VLAs can retain relevant knowledge from prior tasks despite performance degradation during learning new tasks. This knowledge retention enables rapid recovery of seemingly forgotten skills through finetuning. Together, these insight
Stars with masses between roughly 1 and 8~$M_\odot$ end their lives on the asymptotic giant branch (AGB), when intense mass loss takes place. The outflows are generally accepted to be driven by radiation pressure acting on dust grains that form in the dense extended atmospheres created by the action of convection and stellar pulsations. The complex physics underlying convection, stellar pulsations, and dust nucleation precludes predicting AGB mass loss from first principles. We investigated the evolution of two lobes observed to be expanding away from the AGB star Mira~A using images of polarized light obtained at six epochs using SPHERE on the VLT and of molecular emission at two epochs obtained with ALMA. While dust seems confined to the edges of the lobes, gas fills the lobes and displays higher densities than expected at the observed radii based on the large-scale mass-loss rate of Mira~A, with a total gas mass in the lobes of $\sim 2 \times 10^{-5}~M_\odot$. We find the expansion of the lobes to be consistent with both a constant velocity (ejection time in 2010 or 2011) or a decelerating expansion (ejection time in 2012). If ejection events with a similar magnitude happen peri
The principle of rewarding a crowd for surprisingly common answers has been used in the literature for designing a number of truthful information elicitation mechanisms. A related method has also been proposed in the literature for better aggregation of crowd wisdom. Drawing a comparison between crowd based collective intelligence systems and large language models, we define the notion of 'surprisingly likely' textual response of a large language model. This notion is inspired by the surprisingly common principle, but tailored for text in a language model. Using benchmarks such as TruthfulQA and openly available LLMs: GPT-2 and LLaMA-2, we show that the surprisingly likely textual responses of large language models are more accurate in many cases compared to standard baselines. For example, we observe up to 24 percentage points aggregate improvement on TruthfulQA and up to 70 percentage points improvement on individual categories of questions in this benchmark. We also provide further analysis of the results, including the cases when surprisingly likely responses are less or not more accurate.
Reward models are central to aligning LLMs with human preferences, but they are costly to train, requiring large-scale human-labeled preference data and powerful pretrained LLM backbones. Meanwhile, the increasing availability of high-quality synthetic instruction-following datasets raises the question: can simpler, reference-based metrics serve as viable alternatives to reward models during RL-based alignment? In this paper, we show first that BLEU, a basic string-matching metric, surprisingly matches strong reward models in agreement with human preferences on general instruction-following datasets. Based on this insight, we develop BLEUBERI, a method that first identifies challenging instructions and then applies Group Relative Policy Optimization (GRPO) using BLEU directly as the reward function. We demonstrate that BLEUBERI-trained models are competitive with models trained via reward model-guided RL across four challenging instruction-following benchmarks and three different base language models. A human evaluation further supports that the quality of BLEUBERI model outputs is on par with those from reward model-aligned models. Moreover, BLEUBERI models generate outputs that a
A key challenge in evaluating VLMs is testing models' ability to analyze visual content independently from their textual priors. Recent benchmarks such as BLINK probe visual perception through visual prompting, where questions about visual content are paired with coordinates to which the question refers, with the coordinates explicitly marked in the image itself. While these benchmarks are an important part of VLM evaluation, we find that existing models are surprisingly fragile to seemingly irrelevant details of visual prompting: simply changing a visual marker from red to blue can completely change rankings among models on a leaderboard. By evaluating nine commonly-used open- and closed-source VLMs on two visually prompted tasks, we demonstrate how details in benchmark setup, including visual marker design and dataset size, have a significant influence on model performance and leaderboard rankings. These effects can even be exploited to lift weaker models above stronger ones; for instance, slightly increasing the size of the visual marker results in open-source InternVL3-8B ranking alongside or better than much larger proprietary models like Gemini 2.5 Pro. We further show that l
Link prediction is a core challenge in graph machine learning, demanding models that capture rich and complex topological dependencies. While Graph Neural Networks (GNNs) are the standard solution, state-of-the-art pipelines often rely on explicit structural heuristics or memory-intensive node embeddings -- approaches that struggle to generalize or scale to massive graphs. Emerging Graph Transformers (GTs) offer a potential alternative but often incur significant overhead due to complex structural encodings, hindering their applications to large-scale link prediction. We challenge these sophisticated paradigms with PENCIL, an encoder-only plain Transformer that replaces hand-crafted priors with attention over sampled local subgraphs, retaining the scalability and hardware efficiency of standard Transformers. Through experimental and theoretical analysis, we show that PENCIL extracts richer structural signals than GNNs, implicitly generalizing a broad class of heuristics and subgraph-based expressivity. Empirically, PENCIL outperforms heuristic-informed GNNs and is far more parameter-efficient than ID-embedding--based alternatives, while remaining competitive across diverse benchmar
Video generation models have recently achieved impressive visual fidelity and temporal coherence. Yet, they continue to struggle with complex, non-rigid motions, especially when synthesizing humans performing dynamic actions such as sports, dance, etc. Generated videos often exhibit missing or extra limbs, distorted poses, or physically implausible actions. In this work, we propose a remarkably simple reward model, HuDA, to quantify and improve the human motion in generated videos. HuDA integrates human detection confidence for appearance quality, and a temporal prompt alignment score to capture motion realism. We show this simple reward function that leverages off-the-shelf models without any additional training, outperforms specialized models finetuned with manually annotated data. Using HuDA for Group Reward Policy Optimization (GRPO) post-training of video models, we significantly enhance video generation, especially when generating complex human motions, outperforming state-of-the-art models like Wan 2.1, with win-rate of 73%. Finally, we demonstrate that HuDA improves generation quality beyond just humans, for instance, significantly improving generation of animal videos and
Recent theoretical results demonstrate that the convergence rates of permutation-based SGD (e.g., random reshuffling SGD) are faster than uniform-sampling SGD; however, these studies focus mainly on the large epoch regime, where the number of epochs $K$ exceeds the condition number $κ$. In contrast, little is known when $K$ is smaller than $κ$, and it is still a challenging open question whether permutation-based SGD can converge faster in this small epoch regime (Safran and Shamir, 2021). As a step toward understanding this gap, we study the naive deterministic variant, Incremental Gradient Descent (IGD), on smooth and strongly convex functions. Our lower bounds reveal that for the small epoch regime, IGD can exhibit surprisingly slow convergence even when all component functions are strongly convex. Furthermore, when some component functions are allowed to be nonconvex, we prove that the optimality gap of IGD can be significantly worse throughout the small epoch regime. Our analyses reveal that the convergence properties of permutation-based SGD in the small epoch regime may vary drastically depending on the assumptions on component functions. Lastly, we supplement the paper with
When a dataset contains forecasts on unscheduled events, such as natural catastrophes, outcomes may be censored or ``hidden'' since some events have not yet occurred. This article finds that this can lead to a selection bias which affects the perceived accuracy and calibration of forecasts. This selection bias can be eliminated by excluding forecasts on outcomes which have been verified surprisingly early.
An important problem on social information sites is the recovery of ground truth from individual reports when the experts are in the minority. The wisdom of the crowd, i.e. the collective opinion of a group of individuals fails in such a scenario. However, the surprisingly popular (SP) algorithm~\cite{prelec2017solution} can recover the ground truth even when the experts are in the minority, by asking the individuals to report additional prediction reports--their beliefs about the reports of others. Several recent works have extended the surprisingly popular algorithm to an equivalent voting rule (SP-voting) to recover the ground truth ranking over a set of $m$ alternatives. However, we are yet to fully understand when SP-voting can recover the ground truth ranking, and if so, how many samples (votes and predictions) it needs. We answer this question by proposing two rank-order models and analyzing the sample complexity of SP-voting under these models. In particular, we propose concentric mixtures of Mallows and Plackett-Luce models with $G (\ge 2)$ groups. Our models generalize previously proposed concentric mixtures of Mallows models with $2$ groups, and we highlight the importan
Natural language processing has seen rapid progress over the past decade. Due to the speed of developments, some practices get established without proper evaluation. Considering one such case and focusing on reading comprehension, we ask our first research question: 1) How does the order of inputs -- i.e., question and context -- affect model performance? Additionally, given recent advancements in input emphasis, we ask a second research question: 2) Does emphasizing either the question, the context, or both enhance performance? Experimenting with 9 large language models across 3 datasets, we find that presenting the context before the question improves model performance, with an accuracy increase of up to $31\%$. Furthermore, emphasizing the context yields superior results compared to question emphasis, and in general, emphasizing parts of the input is particularly effective for addressing questions that models lack the parametric knowledge to answer. Experimenting with both prompt-based and attention-based emphasis methods, we additionally find that the best method is surprisingly simple: it only requires concatenating a few tokens to the input and results in an accuracy improvem
The mass distribution of neutron stars encodes information about their formation and binary evolution. We compare the masses of two distinct populations: I) the recently identified Gaia neutron stars in wide orbits with solar-like companions and, II) the assumed first-born recycled pulsar in Galactic double neutron star systems. Naively, one would expect their masses to differ due to both the presumed differences in their evolutionary histories, as well as astrophysical selection effects that can filter out configurations that would merge or be disrupted. Yet, we find that their mass distributions are strikingly similar. Using a two-component Gaussian model, we find that both populations exhibit a narrow component centred near $1.3 \text{ M}_\odot$, accompanied by a broader, higher-mass component that extends the distribution toward larger masses. The highest density regions of their fitted parameter posteriors coincide by over 91.6%. Statistical tests further confirm the agreement between these distributions with a Jensen-Shannon divergence $JS < 0.08$ and an earth mover's distance of $W <0.063 \text{ M}_\odot$ at 90% credibility. This finding seems to imply that both mass f
Most large multimodal models (LMMs) are implemented by feeding visual tokens as a sequence into the first layer of a large language model (LLM). The resulting architecture is simple but significantly increases computation and memory costs, as it has to handle a large number of additional tokens in its input layer. This paper presents a new architecture DeepStack for LMMs. Considering $N$ layers in the language and vision transformer of LMMs, we stack the visual tokens into $N$ groups and feed each group to its aligned transformer layer \textit{from bottom to top}. Surprisingly, this simple method greatly enhances the power of LMMs to model interactions among visual tokens across layers but with minimal additional cost. We apply DeepStack to both language and vision transformer in LMMs, and validate the effectiveness of DeepStack LMMs with extensive empirical results. Using the same context length, our DeepStack 7B and 13B parameters surpass their counterparts by \textbf{2.7} and \textbf{2.9} on average across \textbf{9} benchmarks, respectively. Using only one-fifth of the context length, DeepStack rivals closely to the counterparts that use the full context length. These gains are
Leveraging the photometric data of the Sloan Digital Sky Survey and the Galaxy Evolution Explorer (GALEX), we construct mean/median spectral energy distributions (SEDs) for unique bright quasars in redshift bins of 0.2 and up to $z \simeq 3$, after taking the GALEX non-detection into account. Further correcting for the absorption of the intergalactic medium, these mean/median quasar SEDs constitute a surprisingly redshift-independent mean/median composite SED from the rest-frame optical down to $\simeq 500~{\rm \mathring A}$ for quasars with bolometric luminosity brighter than $10^{45.5}~{\rm erg s^{-1}}$. Moreover, the mean/median composite quasar SED is plausibly also independent of black hole mass and Eddington ratio, and suggests similar properties of dust and gas in the quasar host galaxies since cosmic noon. Both the mean and median composite SEDs are nicely consistent with previous mean composite quasar spectra at wavelengths beyond $\simeq 1000~{\rm \mathring A}$, but at shorter wavelengths, are redder, indicating, on average, less ionizing radiation than previously expected. Through comparing the model-predicted to the observed composite quasar SEDs, we favor a simply trun
In the recently introduced topological distance games, strategic agents need to be assigned to a subset of vertices of a topology. In the assignment, the utility of an agent depends on both the agent's inherent utilities for other agents and its distance from them on the topology. We study the computational complexity of finding individually rational outcomes; this notion is widely assumed to be the very minimal stability requirement and requires that the utility of every agent in a solution is non-negative. We perform a comprehensive study of the problem's complexity, and we prove that even in very basic cases, deciding whether an individually rational solution exists is intractable. To reach at least some tractability, one needs to combine multiple restrictions of the input instance, including the number of agents and the topology and the influence of distant agents on the utility.
Performance prediction has been a key part of the neural architecture search (NAS) process, allowing to speed up NAS algorithms by avoiding resource-consuming network training. Although many performance predictors correlate well with ground truth performance, they require training data in the form of trained networks. Recently, zero-cost proxies have been proposed as an efficient method to estimate network performance without any training. However, they are still poorly understood, exhibit biases with network properties, and their performance is limited. Inspired by the drawbacks of zero-cost proxies, we propose neural graph features (GRAF), simple to compute properties of architectural graphs. GRAF offers fast and interpretable performance prediction while outperforming zero-cost proxies and other common encodings. In combination with other zero-cost proxies, GRAF outperforms most existing performance predictors at a fraction of the cost.
The problem of grasping objects using a multi-finger hand has received significant attention in recent years. However, it remains challenging to handle a large number of unfamiliar objects in real and cluttered environments. In this work, we propose a representation that can be effectively mapped to the multi-finger grasp space. Based on this representation, we develop a simple decision model that generates accurate grasp quality scores for different multi-finger grasp poses using only hundreds to thousands of training samples. We demonstrate that our representation performs well on a real robot and achieves a success rate of 78.64% after training with only 500 real-world grasp attempts and 87% with 4500 grasp attempts. Additionally, we achieve a success rate of 84.51% in a dynamic human-robot handover scenario using a multi-finger hand.
Deep reinforcement learning (RL) is notoriously impractical to deploy due to sample inefficiency. Meta-RL directly addresses this sample inefficiency by learning to perform few-shot learning when a distribution of related tasks is available for meta-training. While many specialized meta-RL methods have been proposed, recent work suggests that end-to-end learning in conjunction with an off-the-shelf sequential model, such as a recurrent network, is a surprisingly strong baseline. However, such claims have been controversial due to limited supporting evidence, particularly in the face of prior work establishing precisely the opposite. In this paper, we conduct an empirical investigation. While we likewise find that a recurrent network can achieve strong performance, we demonstrate that the use of hypernetworks is crucial to maximizing their potential. Surprisingly, when combined with hypernetworks, the recurrent baselines that are far simpler than existing specialized methods actually achieve the strongest performance of all methods evaluated. We provide code at https://github.com/jacooba/hyper.
In a recent, strongly emergent literature on few-shot CLIP adaptation, Linear Probe (LP) has been often reported as a weak baseline. This has motivated intensive research building convoluted prompt learning or feature adaptation strategies. In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline, in which the linear classifier weights are learnable functions of the text embedding, with class-wise multipliers blending image and text knowledge. As our objective function depends on two types of variables, i.e., the class visual prototypes and the learnable blending parameters, we propose a computationally efficient block coordinate Majorize-Minimize (MM) descent algorithm. In our full-batch MM optimizer, which we coin LP++, step sizes are implicit, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. By examining the mathematical properties of our loss (e.g., Lipschitz gradient continuity), we build majorizing functions yielding data-driven learning rates and derive approximations of the loss's minima, which provide data-informed initialization of the variables. Ou
Existing approaches to unsupervised video instance segmentation typically rely on motion estimates and experience difficulties tracking small or divergent motions. We present VideoCutLER, a simple method for unsupervised multi-instance video segmentation without using motion-based learning signals like optical flow or training on natural videos. Our key insight is that using high-quality pseudo masks and a simple video synthesis method for model training is surprisingly sufficient to enable the resulting video model to effectively segment and track multiple instances across video frames. We show the first competitive unsupervised learning results on the challenging YouTubeVIS-2019 benchmark, achieving 50.7% APvideo^50 , surpassing the previous state-of-the-art by a large margin. VideoCutLER can also serve as a strong pretrained model for supervised video instance segmentation tasks, exceeding DINO by 15.9% on YouTubeVIS-2019 in terms of APvideo.